What is the most efficient way to initialize a 3D vector?

Question

I have a 3D string vector in C++:

vector>> some_vector

That I am trying is to find a fast method to all

Answer 1

I think I'd optimize it by allocating one large block of memory instead of a lot of little ones. This one is only 2D instead of 3D, but gives the basic idea:

template <class T>
class matrix { 
    size_t columns_;
    std::vector<T> data;
public:
    matrix(size_t columns, size_t rows) : columns_(columns), data(columns*rows) {}

    T &operator()(size_t column, size_t row) { return data[row*columns_+column]; }
};

For 3D, you'll need to deal with "planes" (or something) along with rows and columns, but the basic idea is pretty much the same.

Answer 2

I added several features to Mike Seymour's code such as dynamically resize the 3d vector and on access/assign bounds checking for data vector.

template <typename T>
class vector3d 
{
public:
    vector3d(size_t d1=0, size_t d2=0, size_t d3=0, T const & t=T()) :
        d1(d1), d2(d2), d3(d3), data(d1*d2*d3, t)
    {}

    T & operator()(size_t i, size_t j, size_t k) 
    {
            return (i<=d1 && j<=d2 && k<=d3) ? data[i*d2*d3 + j*d3 + k] 
                                             : data.at(i*d2*d3 + j*d3 + k);
    }

    T const & operator()(size_t i, size_t j, size_t k) const 
    {
        return data[i*d2*d3 + j*d3 + k];
    }

    void resize(const size_t _d1=0, const size_t _d2=0, const size_t _d3=0)
    {
        data.resize(_d1*_d2*_d3);
        d1=_d1;
        d2=_d2;
        d3=_d3;
    }

    void shrink_to_fit()
    {
        data.shrink_to_fit();
    }

    const size_t length() const
    {
        return data.size();
    }

    const size_t capacity() const
    {
        return data.capacity();
    }

    const size_t x() const
    {
        return d1;
    }

    const size_t y() const
    {
        return d2;
    }

    const size_t z() const
    {
        return d3;
    }


private:
    size_t d1,d2,d3;
    std::vector<T> data;
};

Usage:

vector3d<int> vec3d(2,2,2,31); //create 2x2x2 3d vector and fill it with 31
vec3d(1,1,2)=45;               //assign 45 at vec3d(1,1,2)
vec3d.resize(2,2,1);           //resize the vec3d to 2x2x1
vec3d(1,2,2)=67;               //error (its out of bounds)

Answer 3

May I know what cause such difference?

The first version constructs a 2-d vector by copying a 1-d vector, and then constructs the 3-d vector by copying that. This might be slower than resizing the vectors without copying. However, I'd hope that the difference would be negligible if you're building with optimisation.

And is there better way to allocate memory for a 3D vector?

It might be better to use a single contiguous array, wrapped in a class that provides multi-dimensional accessors. This would make allocation much simpler, and would also avoid some pointer dereferencing when accessing elements (at the cost of a bit of arithmetic). Something like this:

template <typename T>
class vector3d {
public:
    vector3d(size_t d1=0, size_t d2=0, size_t d3=0, T const & t=T()) :
        d1(d1), d2(d2), d3(d3), data(d1*d2*d3, t)
    {}

    T & operator()(size_t i, size_t j, size_t k) {
        return data[i*d2*d3 + j*d3 + k];
    }

    T const & operator()(size_t i, size_t j, size_t k) const {
        return data[i*d2*d3 + j*d3 + k];
    }

private:
    size_t d1,d2,d3;
    std::vector<T> data;
};

Answer 4

To initialize a 3D string vector you shall initialize the vector structure for each dimension one at a time and for each index, for instance:

  vector<vector<vector<string> > > myvector; //declare the 3D vector
  for(k=0; k<3; k++)
  {
    myvector.push_back(vector<vector<string> >()); //initialize the first index with a 2D vector
    for(i=0; i<4; i++)
    {
      myvector[k].push_back(vector<string>()); //initialize the 2 index with a row of strings
      for(j=0; j<4; j++)
      {
         result = " whatever you want to insert in the vector element";
         myvector[k][i].push_back(result); //fulfill the last index regularly
      } 
    }
  }

Answer 5

When you initialize a vector of vectors in the first method, a temporary vector is allocated and then copied into the outer vector the required number of times. This means you have an extra allocation that is unnecessary and the new elements are initialized by copying their value from another data structure, which uses more memory accesses.

Resizing the vectors as per the second method is more ugly but avoids the extra allocation. Furthermore the new elements are created by the default constructor and do not need to copy from other vectors. This will also be faster.

If speed matters (and maybe it doesn't, premature optimization and all that), then you must use the second method (OR a single-block allocation as suggested by the other answers). I don't have faith that a compiler can simply "optimize" away the inefficiency of the first method.